Guide for computing professionals: what to study when AI compresses technology
Where the profession came from, what AI has already changed in the job market and in research, what the new computing professional looks like, and a study path that goes beyond code: enterprise architecture, scientific method, philosophy, science, business, economics and languages.
- 7 sections
- 10 interactions
- 5 study stages
- 8 disciplines beyond technology
Each generation of computing turned the previous generation’s work into a ready-made layer
The history of the profession can be told as a sequence of compressions. What once required a specialist and weeks of work becomes, a few years later, a tool, a library or a service that anyone in the field uses in minutes. The compiler compressed machine-language programming. The relational database compressed file management. The cloud compressed buying and running servers. Generative AI is compressing the writing of code.
None of these shifts ended the profession. What happened was a displacement: the repetitive work moved down into the new layer, and the value moved up to what the layer cannot solve on its own. Choosing the right problem, specifying it precisely, integrating the parts, verifying the result and answering for the consequences. Understanding this pattern is the first step in deciding what to study now.
The professional had to know the machine inside out
The ENIAC programmers configured the computer with cables and switches, working from the machine’s diagrams. There was no language and no operating system between the person and the hardware. Everything was specialist work.
FORTRAN and high-level languages
IBM releases FORTRAN, and the compiler starts translating formulas written by people into machine instructions. Many doubted that an automatically generated program could be as efficient as a hand-written one. The doubt lost, and the number of people able to program multiplied.
The “software crisis” and the birth of Software Engineering
A conference sponsored by NATO establishes the term “software engineering”. The diagnosis: systems were growing faster than the ability to build them on time, on budget and with quality. Programming well was no longer enough; method was needed. [1]
Edgar Codd and the relational model
Codd’s paper separates the logical form of data from the way it is stored on disk. Ten years later, relational databases take file management away from every program, and data modeling becomes a core skill.
John Zachman and enterprise architecture
In the IBM Systems Journal, Zachman proposes looking at information systems as part of the architecture of the whole organization: who decides, which processes exist, which data flows, which technology supports it all. In 1995 the first version of TOGAF appears, today the most widely used standard in the field. [2] [12]
The Agile Manifesto
Seventeen practitioners sign the manifesto that prioritizes working software, customer collaboration and responding to change. The professional starts working closely with the business, in short cycles.
Public cloud computing
Amazon Web Services launches S3 storage and EC2 virtual machines. Buying, installing and running servers, weeks of work, becomes an API call. The question changes from “do we have a machine?” to “how much does it cost, how much can it take and what happens when it fails?”.
DevOps: whoever builds it also runs it
The first DevOpsDays gives a name to a movement that brings development and operations together. Continuous delivery, infrastructure as code and observability become part of the job of whoever writes the system.
The Transformer architecture
The paper “Attention Is All You Need” introduces the neural network architecture behind today’s large language models. AI research starts scaling with data and compute as never before.
Generative AI and agents
Assistants based on large language models start writing, explaining and reviewing code, and then carrying out multi-step tasks inside repositories. Generative AI reached 53% adoption in three years, faster than the personal computer and the internet at the same stage. [9] This is the compression behind this guide.
The same task in three eras
A concrete example of the shift: putting a web system online for a company’s customers. Compare what was manual work in each era and what remained the professional’s responsibility.
- The manual work
- Buy or rent a physical server, install the operating system, configure the network, database and web server, write all the code and copy the files to the machine.
- Typical time
- Weeks, sometimes months, just for the infrastructure.
- What was scarce
- People who could set up and maintain the machine.
- What remained yours
- Understanding the customer’s problem, deciding what to build, making sure it worked.
- The manual work
- Create virtual machines in the cloud in minutes, automate configuration with scripts, use frameworks that already handle authentication, data access and pages.
- Typical time
- Days for the infrastructure; the effort moves to the product code.
- What was scarce
- People who could design systems that scale and run them in production.
- What remained yours
- Understanding the customer’s problem, deciding what to build, making sure it worked.
- The manual work
- Describe what you want to AI assistants that generate the code, the tests, the infrastructure as code and the delivery pipeline, on top of managed platforms.
- Typical time
- Hours to have something running. The time left over goes to reviewing, integrating and verifying.
- What is scarce
- People who can tell whether what was generated is correct, secure, sustainable and aligned with the business.
- What remains yours
- Understanding the customer’s problem, deciding what to build, making sure it works. Now with less time spent on everything else, and more exposure when this part fails.
The last line does not change across eras. It is the core of the profession, and it is what your study should aim at.
This idea, that each layer of abstraction changes where the judgment of the people working with it is formed, is the subject of the essay A camada que ensina: compressão, abstração e a formação do julgamento em engenharia de software (the layer that teaches), published on the blog, under Essays (in Portuguese).
What the data shows about work in computing after generative AI
Adoption is practically universal among people who build software. The gains are not. The most solid studies so far point to a market that pays less and less for isolated execution and more and more for people who can specify, verify and integrate.
Four changes we can already state
For those already in the field, the practical reading is this: the work that survives compression is the work that requires context. Context about the business, about the architecture that already exists, about cost and regulatory constraints, about the people who will use the system. That context is not in the code, which is why AI does not generate it on its own.
When writing code gets cheap, the expensive part becomes knowing what to write, where it fits and how to prove it is right.
Computing research has become more empirical, more expensive and more intertwined with other sciences
AI did not change only the job market. It changed who does frontier research, how a result is evaluated and where computing meets other fields of knowledge.
The model frontier left the university
According to Stanford University’s AI Index 2025, nearly 90% of the AI models considered notable in 2024 came from industry, up from 60% the year before. Training a frontier model requires data and compute that few academic labs have. [8]
What this asks of you: reading and evaluating research done outside academia, with less transparency about data and methods, and telling a result from an announcement.
Measuring became a research problem in itself
Models quickly hit the ceiling of benchmarks, and a good benchmark score does not always translate into real-world gains, as METR’s experiment with developers showed. Rigorous evaluation, with a control group, a sample and a stated hypothesis, is back at the center of the field. [6]
What this asks of you: statistics, experimental design and the discipline not to mistake impression for evidence.
Computing became an instrument of the other sciences
In 2024, the Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton, for work underpinning neural networks, and the Chemistry prize was shared by David Baker and the creators of AlphaFold, Demis Hassabis and John Jumper, for protein structure prediction. [10]
What this asks of you: knowing enough of another science to talk with the people who work in it. The most interesting problems are at the boundaries.
Classic topics gained practical urgency
Formal verification, security, reliability of distributed systems, data quality and governance, human-computer interaction and ethics are no longer niche subjects. When part of the code is generated by a machine, guaranteeing system properties becomes more important, not less.
What this asks of you: fundamentals. They are what let you ask whether the system does what it should, not just whether it compiles.
The reference curriculum has already absorbed the change
CS2023, the reference curriculum from ACM, the IEEE Computer Society and, for the first time, AAAI, organizes Computer Science into 17 knowledge areas and treats Society, Ethics and the Profession as an integral part of most of them, not as a separate course. [11]
What this asks of you: if your education predates this change, your most likely gaps are in AI, security and the social consequences of what gets built.
Technical depth, an architectural view and a repertoire in other disciplines
The new computing professional is not a generalist who knows a little of everything. They remain deep in computing, because that depth is what makes it possible to verify what AI produces. The difference is that they also see how the system fits into the organization and have a repertoire in other fields to converse, decide and answer for their choices.
A useful image is the Greek letter π. The top bar is breadth: method, business, economics, science, philosophy, languages. The two legs are the depths: computing fundamentals and architecture, from software to the enterprise. With only one leg, a person can execute but struggles to decide. Without the legs, they talk well and deliver little.
The four dimensions of the profile
Eight statements about your current work. Answer based on what you do, not on what you would like to do. The result is not stored anywhere.
I can explain why AI-generated code is right or wrong, not just whether it passes the tests.
I can draw how the system I work on connects to the company’s other systems.
I can explain how my work affects the organization’s revenue, cost or risk.
I know how much the system I help maintain costs per month, and how much one hour of downtime costs.
I read research papers in my area and can judge whether the method supports the conclusion.
Before claiming something improved, I measure it, with a metric defined in advance.
I write decision documents that people from other areas can understand.
I can negotiate and defend a technical decision in English.
Answer all 8 statements to see the result.
Enterprise architecture is the discipline that links an organization’s strategy to the systems that support it
Enterprise architecture is the study of how an organization structures itself to carry out its strategy: which capabilities it needs, which processes deliver those capabilities, which data they use, which applications support them and which technology runs it all. The best-known frameworks are Zachman’s, from 1987, and TOGAF, maintained by The Open Group and now in its tenth edition. [2] [12]
For a long time, it seemed like a topic for large corporations and documents nobody read. With AI, it is back at the center for a practical reason: a useful AI agent needs access to the right data, must respect permissions, has to fit into a business process and must be auditable. The model is the easy part; the hard part is the fit. It is the same finding as DORA 2025, seen from the organization’s side: AI pays off where the architecture allows it. [4]
Click each domain to see the question it answers and an AI example.
Which capabilities does the organization need to carry out its strategy?
Maps capabilities (selling, serving, billing, complying), processes, roles and indicators. It is the layer that gives meaning to all the others.
With AI: before choosing a model, decide which capability it improves, which process it enters and who answers for the result.
What information exists, who owns it and how far can it be trusted?
Data models, lineage, quality, ownership and access rules. In almost every AI project, this is where the schedule blows up.
With AI: an assistant that answers questions about contracts is only reliable if the contract base has an owner, versions and access control.
Which systems exist, what does each one do and how do they talk to each other?
The application portfolio, integrations, APIs, what is bought and what is built, what should be retired.
With AI: an agent that opens tickets, looks up orders and reissues invoices needs stable APIs on the right systems, not direct access to the database.
Where does everything run, how much does it cost and what happens when it fails?
Cloud, networks, platforms, security, observability and capacity. It is the layer most compressed over the last two decades, and still the one that punishes wrong decisions the most.
With AI: estimating inference usage and cost, acceptable latency and the plan for when the model provider goes down.
Who decides, under which rules, and how is it verified?
Architecture principles, standards, compliance with laws such as Brazil’s LGPD, risk management and auditing.
With AI: recording which model made which decision, with which data, and allowing a person to review it. The research on the Data-Centric Trust Pipeline addresses this problem in sensitive domains.
This is also the focus of the workshops on Cloud, Data and AI as a journey of enterprise architecture transformation, described in Talks & Workshops.
A five-stage study path, with technology and other disciplines side by side
The stages follow an order, but nobody starts from zero: use the self-assessment in section 04 to decide where to come in. Each stage lists what to study in technology, what to study outside it, reference readings and a deliverable that proves the learning. A suggested pace is four to six hours a week, for three to six months per stage.
Recover the fundamentals AI does not replace. They are the basis for verifying anything a tool produces.
- Algorithms and complexity: estimating the cost of a solution before running it
- Data structures and how they behave in memory
- Operating systems, concurrency and networks
- Databases: modeling, transactions and consistency
- Testing and verification: what a test proves and what it doesn’t
- Logic and discrete mathematics, revisited with adult eyes
- Basic probability and statistics
- Problem solving as a method, not a talent
- Technical writing: explaining precisely in one paragraph
Move from code to the system. Understand how the parts fit together, fail and cost.
- Architecture patterns and their costs: monolith, microservices, events, serverless
- Distributed systems: consistency, partitioning, queues, idempotency
- Cloud computing, capacity and observability
- Security by design
- Architecture decision records
- Systems thinking: feedback, delays and second-order effects
- Basic microeconomics: opportunity cost and trade-offs
- Finance basics: fixed cost, variable cost, investment
Move from the system to the organization. See the company as a set of capabilities supported by data, applications and technology.
- The domains of enterprise architecture and the TOGAF cycle
- Capability maps and application portfolio
- Data architecture and governance
- Integration between systems and API strategy
- Domain-driven design to draw boundaries
- Strategy: how companies compete and what sets them apart
- Business processes and indicators
- Accounting and finance for non-specialists
- Basics of digital law, starting with data protection
Treat AI as a system component, not as magic. Know how to choose, integrate, evaluate and govern it.
- Machine learning and how large language models work
- Engineering systems with AI: context retrieval, tools, agents
- Evaluating models with your own test sets
- Inference cost, latency and capacity
- AI-specific security: prompt injection and data leakage
- Applied statistics and experimental design
- Ethics and philosophy: responsibility, bias, explainability
- Regulation: data protection laws and the EU AI Act
- Psychology of decision-making: overconfidence and automation bias
Produce knowledge and decisions that other people follow. This is where breadth becomes influence.
- Delivery and reliability metrics, such as DORA’s
- Evaluating emerging technology: when to adopt, when to wait
- Empirical studies in software engineering
- Cross-team architecture review
- Research methodology: question, hypothesis, method, validity
- Philosophy of science: what counts as evidence
- Economics of innovation and digital platforms
- Communication: writing, presenting and teaching, in more than one language
What to study outside computing, and why each field makes a difference at work
None of these fields requires a second degree. They require enough to think with their tools. Click each one to see why it matters, how it shows up at work and where to start.
To separate what we know from what we think
Asking a question that can be answered, choosing a method, defining the metric before measuring and recognizing the limits of the result. It is the antidote to decisions made on impressions, including the impression that a tool made the team faster.
At work: A/B tests, technology pilots, model evaluation, incident analysis.
Where to start: Wazlawick, Metodologia de Pesquisa para Ciência da Computação (in Portuguese).
To think rigorously about knowledge, language and responsibility
Logic is at the base of computing. Epistemology asks what counts as evidence. Ethics asks who answers for what a system does. Philosophy of mind and of language helps avoid overstating or understating what AI models do.
At work: discussing bias, explainability, agent autonomy and acceptable use of data.
Where to start: Thomas Nagel, What Does It All Mean?; Cathy O’Neil, Weapons of Math Destruction.
To work where computing meets other fields
Physics teaches modeling and orders of magnitude. Statistics teaches uncertainty. Biology, climate, health and finance are where some of AI’s hardest and best-funded problems are. Speaking the language of another science opens doors technique alone does not.
At work: projects in health, energy, agribusiness, financial markets, industry.
Where to start: Richard Feynman, The Character of Physical Law; an introductory statistics course.
To know why the system exists
Strategy, business model, product, sales and operations. Whoever understands how the company makes money proposes the right solution before anyone asks, and takes part in decisions instead of just receiving them.
At work: prioritizing the backlog, defending an investment in architecture, designing a product.
Where to start: Michael Porter, “What Is Strategy?”; Clayton Christensen, The Innovator’s Dilemma.
To understand cost, incentives and scale
Opportunity cost, incentives, network effects and marginal cost explain why some platforms dominate markets, why the cloud charges the way it does and how the economics of AI, with per-use model costs, changes a product’s numbers.
At work: estimating the cost of an AI feature, choosing between build and buy, negotiating with vendors.
Where to start: N. Gregory Mankiw, Principles of Economics; Carl Shapiro and Hal Varian, Information Rules.
To be understood, at home and abroad
English is the language of documentation, research and many remote jobs. Spanish opens Latin America. And writing well in your own language is still rare: specifications, proposals and decisions are texts, and now they are also instructions for AI.
At work: writing specifications, running meetings with clients abroad, publishing, giving talks.
Where to start: William Zinsser, On Writing Well; one technical presentation a month in a second language.
To build what people can and want to use
Design, cognitive psychology and human-computer interaction explain why a correct system can fail and why people trust automated answers too much. With conversational interfaces, this has become even more important.
At work: designing AI flows in which people know when to review, accessibility, adoption of internal systems.
Where to start: Don Norman, The Design of Everyday Things.
To know what can be done, and with what care
Data protection, intellectual property, software contracts and AI regulation set design limits. In Brazil, the reference is the General Data Protection Law (LGPD, Law No. 13,709/2018); in the European Union, the AI Act (Regulation (EU) 2024/1689) already affects companies that serve European customers.
At work: designing systems that process personal data, documenting automated decisions, responding to audits.
Where to start: the text of the data protection law that applies to you and the guidance from its authority.
Habits that sustain the path
Content without habit does not accumulate. Eight habits. Click each card to see why the habit matters and how to start.
Myth or truth?
Statements that circulate among professionals in the field. Answer each one and see what the data says.
“With AI writing code, computing fundamentals have lost their value.”
The more code is generated, the more it matters who can verify it. Complexity, concurrency, security and consistency are exactly what lets you say whether code that looks right is right.
“People starting their careers are the ones who feel AI’s effect on jobs the most.”
US data show a 13% relative decline in employment for 22- to 25-year-olds in the most exposed occupations, with more experienced workers holding steady. The caveat: these are data from one country, over a short period, and people who use AI to augment their own work did better. [5]
“Using AI always makes an experienced professional faster.”
In METR’s experiment, experienced developers took 19% longer with AI on their own projects, while believing they had gained speed. The effect depends on the task, the context and how the tool is used. [6]
“Being multidisciplinary means knowing a little of everything.”
Breadth without depth turns into talk without delivery. The profile in demand keeps depth in computing and architecture and uses the other disciplines to decide better.
“Enterprise architecture is bureaucracy: diagrams nobody reads.”
Done as document production, it does become bureaucracy. Done as a decision discipline, it is what defines where AI can come in, with which data and under which rules. The problem usually lies in the practice, not in the discipline.
“Reading research papers is only useful for people in academia.”
In AI, almost everything appears first in a paper, often from company labs. Knowing how to read a method and spot its limitations protects you from adopting something that only worked on a benchmark.
“A technical professional doesn’t need to understand finance.”
Every architecture decision has a cost: cloud, inference, maintenance, downtime. Whoever can translate these choices into numbers takes part in the decision; whoever can’t receives it ready-made.
“To work with international clients, reading technical English is enough.”
Reading documentation is the minimum. International work requires conversing, negotiating deadlines, writing proposals and defending decisions in real time. And in Latin America, Spanish opens doors English doesn’t.
The questions that come up most
I’ve been working in software development for ten years. Where do I start?
Take the self-assessment in section 04 and start with the dimension where you scored lowest. For most experienced professionals, the gap is not in code; it is in enterprise architecture, finance or method. Stage 3 of the path is usually the best entry point.
How can I gain enterprise architecture experience without the job title?
Start with the system you work on: map what it talks to, which business capability it supports and which data it uses. Offer that map to the team and to leadership. Architecture is a practice before it is a job title.
Are certifications worth it?
They are worth it as a study roadmap and as a signal on your résumé, especially in cloud and enterprise architecture. They don’t replace the deliverable that proves the learning at each stage: that is what shows you can use what you studied.
Do I need a master’s degree?
Not for most careers. But a master’s, especially a professional one, is the most structured way to learn the scientific method, and it is very valuable for anyone who wants to lead technical decisions, evaluate emerging technologies or teach. The Advising page explains what that journey looks like.
How can I use AI to study without hurting the learning?
Use AI after you try, not before. Ask it to explain your mistake, generate exercises, critique your solution or play the role of an interviewer. Avoid asking for the ready-made answer to something you are trying to learn.
How long does the whole path take?
At the suggested pace of four to six hours a week, each stage takes three to six months. The whole path takes a year and a half to three years, and it never ends: with each new compression of technology, some stages need to be revisited.
Eight actions to set the path in motion
Check what you have done. Your progress is saved only in this browser.
0 of 8 done
Send your question about career, study or architecture.
Use the contact form, or read the technical posts and essays on the blog.
- Naur, P.; Randell, B. (eds.). Software Engineering: Report on a conference sponsored by the NATO Science Committee, Garmisch, October 7–11, 1968. Brussels: NATO, 1969. homepages.cs.ncl.ac.uk
- Zachman, J. A. A Framework for Information Systems Architecture. IBM Systems Journal, v. 26, n. 3, 1987.
- Stack Overflow. 2025 Developer Survey, AI section. survey.stackoverflow.co
- DORA / Google Cloud. State of AI-assisted Software Development, 2025. dora.dev
- Brynjolfsson, E.; Chandar, B.; Chen, R. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab, 2025. digitaleconomy.stanford.edu
- Becker, J.; Rush, N.; Barnes, B.; Rein, D. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR, 2025. arxiv.org
- World Economic Forum. The Future of Jobs Report 2025. weforum.org
- Stanford HAI. AI Index Report 2025. hai.stanford.edu
- Stanford HAI. AI Index Report 2026. hai.stanford.edu
- Nobel Foundation. 2024 Nobel Prizes in Physics and Chemistry. Physics · Chemistry
- ACM, IEEE Computer Society and AAAI. Computer Science Curricula 2023 (CS2023). csed.acm.org
- The Open Group. The TOGAF Standard, 10th edition. opengroup.org
- Ross, J. W.; Weill, P.; Robertson, D. C. Enterprise Architecture as Strategy: Creating a Foundation for Business Execution. Boston: Harvard Business School Press, 2006.
- Wazlawick, R. S. Metodologia de Pesquisa para Ciência da Computação. Rio de Janeiro: Elsevier. In Portuguese.